NVIDIA’s latest creation is the DGX Spark, a tiny AI mini PC designed for scale-out clustering. Priced at $3999, it may seem expensive, but its compact size and potential for massive computing power make it an exciting addition to local and portable AI development.
The DGX Spark features 20 Arm-based CPU cores from Arm’s Cortex-X925 and Cortex-A725 series, paired with a NVIDIA Blackwell GPU. This combination is linked by the C2C interconnect and provides up to 273GB/s of shared memory. What’s more, this device can be used in clusters, making it an attractive option for those who want to create massive computing systems.
The DGX Spark is compact, measuring just over 1L and weighing around 1.2kg, with a stylish design reminiscent of the NVIDIA DGX-1. It has display outputs and four USB4 ports, along with a HDMI port and a 10GbE port. The device also supports 200GbE clustering with a second unit.
NVIDIA is selling two DGX Spark devices as part of a cluster configuration, which includes a QSFP cable to support the connection. The company is initially focusing on bringing out 2x GB10 cluster configurations using the 200GbE RDMA networking.
The device comes pre-installed with the NVIDIA DGX OS, an Ubuntu Linux base with many NVIDIA drivers and goodies baked in. This makes it easy to scale-out AI development with NCCL.
In comparison to other devices like the Apple M3 Ultra studio, the DGX Spark offers a unique combination of scale-out networking capabilities at a lower price point. It’s expected to ship this summer, and its potential for massive computing power is making it an exciting addition to local and portable AI development.
The device is powered by USB-C input, which may seem surprising given its compact size, but provides up to 170W of power consumption.
Source: https://www.servethehome.com/the-nvidia-dgx-spark-is-a-tiny-128gb-ai-mini-pc-made-for-scale-out-clustering-arm